Transformers: Question Answering Modeling through Hugging Face Models

[AishwaryaVerma]

❓ Questions & Help

Details

Hello everyone, I have large number of documents and I need to extract specific information through Hugging Face Question Answering Model. First issue I faced was the document size was very large so it gave me token error and afterwards, I divided the data into small paragraphs, then I applied the given model. But this time, answer was not accurate. So, I just want to know, is there any alternative method or model to do this.

A link to original question on Stack Overflow:

[patil-suraj]

How long is your document, you may wanna try longformer model which can handle sequences upto 4096 tokens. Here's a longformer model trained for QA https://huggingface.co/valhalla/longformer-base-4096-finetuned-squadv1

Also, take a loot at this https://github.com/deepset-ai/haystack. This might help you a lot

[AishwaryaVerma]

How long is your document, you may wanna try longformer model which can handle sequences upto 4096 tokens. Here's a longformer model trained for QA https://huggingface.co/valhalla/longformer-base-4096-finetuned-squadv1

Also, take a loot at this https://github.com/deepset-ai/haystack. This might help you a lot

I looked into it. That is a great help. Thanks. Can we also decide the output length with these type of pretrained models?

[patil-suraj]

Theses QA models aren't generative. So there's no output length constraint

[patrickvonplaten]

@AishwaryaVerma - For QA the output length is usually very small (only a couple of words). It is very rare that the answer of AutoModelForQuestionAnswering is longer than 3,4 words.

You might also want to take a look at: https://huggingface.co/allenai/longformer-large-4096-finetuned-triviaqa

[patrickvonplaten]

And this notebook: https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing